FT称人形机器人短期商业价值有限，更多价值或在具身AI数据而非表演场景

Humanoid robots will be useful, just not as we imagined
Humanoid robots can do astonishing things. As demonstrated at the World Humanoid Robot Games in Beijing this week, they can run, box, play football, take on obstacle courses and perform synchronised cheerleading routines. One robot even beat Usain Bolt’s record for the 100-metre sprint (although it couldn’t slow down in time to avoid smashing into a padded wall afterwards).
Such feats helped Unitree Robotics, a leading Chinese humanoid robot developer, pull off a heavily oversubscribed listing on the Shanghai Stock Exchange this month. Unitree’s shares initially surged, giving it a market value of $51bn at the end of first-day trading, but have fallen back sharply since.
In the US, Elon Musk has been making bombastic claims about Tesla’s Optimus robot, which is increasingly a main focus of his electric car company. This humanoid robot, Musk predicted, would probably be the “biggest product ever”, forecasting up to $10tn in sales in the long term.
At the Beijing games, billed as the “best of the bots”, humanoid robots are the stars and certainly generate a wow factor on YouTube. But as with some human influencers, you wonder if they have much purpose beyond entertaining us. If technology, as the saying goes, is about making machines act like they do in the movies, then these robots certainly fit the bill. But they are more likely to be the show ponies than the workhorses of our robotic future.
Humanoid robots at the games struggled with more dexterous tasks, such as clearing a dining table or plugging in an electric vehicle charger. Besides, some roboticists have always argued that mimicking the human form is an odd way to build the most capable robots. A humanoid robot may have to incorporate 30 axes of motion when simpler designs can perform similar functions with half that number. That can make humanoids costlier and clumsier.
The vast majority of the world’s installed non-humanoid robots are more functional in design and mostly perform single operations, making them a lot less videogenic. These robots do some of the dull, dirty, delicate and dangerous work that humans no longer want to do. They can assemble cars, fetch items in warehouses and perform routine surgery. The world’s militaries are also rushing to deploy drones at massive scale given their devastating use in the Russia-Ukraine war. These are the areas where most robots are currently used and where most money is probably to be made.
Right now, China dominates industrial robotics. The International Federation of Robotics estimates it has an installed base of about 2mn industrial robots, 4.5 times more than Japan, the world’s second-biggest market.
Robotics is also central to China’s 15th Five-Year Plan. As of 2024, China accounted for about 54 per cent of new global installations of industrial robots, with Europe at 16 per cent and the US on 6 per cent.
Like other Asian countries, including Japan and South Korea, China has a strong incentive to adopt robots to compensate for an ageing workforce. Partly as a result, perceptions of robots are far more favourable in Asia than in the US or Europe, easing the path to adoption.
As the world’s software superpower, the US still provides the best robotic “brains.” And the prize for creating the standard robotic operating system of the future may be enormous. The Trump administration is trying to boost domestic robotics manufacturing by restricting mostly Chinese imports. The US Federal Communications Commission has just blocked the import of foreign-made humanoid and quadruped robots, citing national security risks.
That points to one way humanoid robots may prove their economic value: as providers of data to guide future developments. The internet provides vast amounts of information to train AI large language models, such as ChatGPT and Claude. But roboticists desperately need more visual and sensory data to improve how their creations interpret and respond to the physical world. “One of the huge challenges the robotics industry is facing is the lack of training data for AI,” Susanne Bieller, general secretary of the IFR, tells me.
Robotics companies are generating synthetic data for training purposes, most notably in the autonomous car industry. Spookily, some AI companies are also amassing so-called egocentric data, harvested by attaching video cameras to the heads of Indian textile workers to capture their movements, for example. This development raises very obvious concerns about privacy, security and control. But if these can be overcome, humanoid robots, operating alongside humans as collaborative robots (or cobots), could act as interconnectors between AI models and the real world. The ultimate vision is communities of cobots that constantly learn from each other by tapping into a collective “hive” mind.
This school of thought suggests “embodied” AI-powered robots may yet emerge as the more promising path to artificial general intelligence, when AI matches humans across all cognitive tasks. Simply scaling LLMs is not enough. By training robots to perceive, predict, act, fail and adapt, they will develop better causal reasoning and greater autonomy. If that ever happens, humanoid robots may finally earn their star billing, so long as they have learnt not to crash by then.